Lightweight Feature Enhancement Network for Single-Shot Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Lightweight one-stage object detectors offer real-time performance but struggle with detection accuracy.
• Enhancing feature discriminability and robustness is crucial for improving object detection, particularly for small objects.

Purpose of the Study:

• To develop a real-time, high-precision, and lightweight object detection system.
• To improve the detection performance of lightweight models for small objects.

Main Methods:

• Proposed the adaptive receptive field fusion (ARFF) module to enhance feature representation by adaptively learning fusion weights.
• Introduced an enhanced up-sampling (EU) module to minimize information loss during feature map up-sampling.
• Integrated ARFF and EU modules into the YOLO v3 architecture, creating the ARFF-EU network.

Main Results:

• The ARFF-EU network achieved state-of-the-art speed and accuracy trade-offs on Pascal VOC and MS COCO datasets.
• Reported 83.6% AP at 37.5 FPS on Pascal VOC and 42.5% AP at 33.7 FPS on MS COCO.
• Demonstrated significant improvements in detection performance for small objects.

Conclusions:

• The proposed ARFF and EU modules effectively enhance the detection capabilities of lightweight object detectors.
• The ARFF-EU network represents an advancement in developing high-performance, real-time object detection systems.